[comp.graphics] Extracting objects from bitmaps

joshi@wuche2.wustl.edu (Amol Joshi) (03/03/90)

i am currently investigating prior research done in the
area of extracting meaningful objects from bitmaps. is there
any PD software (prototype is just fine) that allows one to
extract simple geometric objects like line, rectangles from
bitmaps? i know that a company by name `CAD Research' offers 
a service to convert blue-prints into AUTOCAD format. they
claim that they use sophisticated AI techniques to get almost
100% accuracy.

well, i would be happy to if i can get any info on:
	1. research papers published in this area or
	2. any PD prototypical systems or
	3. name/address of any commercial outfit  other than the
	   one mentioned above that would offer such services.

i don't know if this issue is of general interest and/or has
been discussed in this newsgroup recently (yes, i am one of
those recent subscribers!). if it has indeed been discussed,
please email your replies. 

thanx a bunch...

:amol
-- 
------------------------------------------------------------
Amol Joshi                         | joshi@wuche2.wustl.edu
Department of Chemical Engineering |
Washington University in St. Louis.|

larry@csccat.UUCP (Larry Spence) (03/03/90)

In article <1990Mar2.215036.28915@wuche2.wustl.edu> joshi@wuche2.wustl.edu () writes:
>
>i am currently investigating prior research done in the
>area of extracting meaningful objects from bitmaps. is there
>any PD software (prototype is just fine) that allows one to
>extract simple geometric objects like line, rectangles from
>bitmaps? i know that a company by name `CAD Research' offers 
>a service to convert blue-prints into AUTOCAD format. they
>claim that they use sophisticated AI techniques to get almost
>100% accuracy.      ^^^^^^^^^^^^^^^^^^^^^^^^^^^        ^^^^^^
 ^^^^^^^^^^^^^

Boy, that's a bunch of marketing-speak, if I may say so.  "Almost
100% accuracy."  Accuracy with respect to what?  An infinite number
of geometric objects can be rasterized to the same bits, so how
does it know WHICH set of objects resulted in the bitmap?  Also,
say you scan in a circle.  The bitmap image of the circle will have
flat spots at multiples of 45 and 90 degrees (at least).  How does
the software "know" that these are artifacts to be ignored, as
opposed to straight line segments?  It's impossible to tell by
just inspecting the bitmap, isn't it?

Autodesk had a program called CadCamera a few years ago that purported
to do the job; they pulled it off the market.  I don't know why, but
I can say that raster->vector conversion is fraught with more unreal-
istic expectations than almost anything I can think of.  "Well, of
course that's a rectangle overlapping another rectangle overlapping
a bunch of circles -- it should be EASY for the program to see that!"
AI can definitely help, but casual claims of "almost 100% accuracy"
should be examined with skepticism.

I think you need to define the problem more precisely (monochrome
bitmaps? greyscale? color? is your bitmap noisy? what is "accuracy"?) 
before you can start working on a solution.

-- 
Larry Spence
larry@csccat
...{texbell,texsun}!csccat!larry